[Beowulf] Re: The move to gigabit - technical questions

[Vincent Diepeveen]

G'morning 

At 01:42 PM 3/17/2005 +1100, steve_heaton at ozemail.com.au wrote:
>G'day all
>
>Somewhat relevant... Part of my benchtesting exersize on my DIY beowulf
was a comparison between running the onboard FastEthernet v's adding an
Intel 1000MT GigaE adapter. 
>
>I changed the MPI config to ensure the MPI traffic had the GigaE to itself
and "other" traffic went via the FastE.
>
>I ran the full MPI perftest suite. Sample graphs on this web page:
>
>http://members.ozemail.com.au/~sheaton/lss/
>-> Computing
>
>It was indeed "a bit" faster.

You ship 1.01E+02 = 1.01 * 10^2 = 101 bytes per block and reach roughly 240
megabit per second with it.

Interesting to know is how much SYSTEM time you lose while shipping 240
megabit per second in say 2MB blocks, not 101 bytes. 101 bytes is real real
little IMHO.

A system is obviously pretty near useless when you have zero processortime
left thanks to network.

>There's some NetPipe results in there too. I can provide more details if
any one is interested.

How much processortime is left while running netpipe.

Do your cards use DMA?

>Note: I know magic can be worked with the Intel driver but 
>this is "vanilla" ATM.

Vanilla is what we need with respect to gigabit.

The non vanilla theoretic 'raw data throughput' and raw latency without
protocol overhead, that's for the highends who for sure have solved the
processor time issue already long long ago :)

>Cheers
>Steve
>
>
>This message was sent through MyMail http://www.mymail.com.au
>
>
>_______________________________________________
>Beowulf mailing list, Beowulf at beowulf.org
>To change your subscription (digest mode or unsubscribe) visit
http://www.beowulf.org/mailman/listinfo/beowulf
>
>